Notion Link Triage

Notion Link Triage

Auto-titles and classifies untitled links in the user's Notion link-inbox database, then surfaces high-signal entries as Linear-issue suggestions.

When to use

• User says "triage my notion links", "process my notion inbox", "inbox-zero notion"
• User pastes a Notion database URL (?v= in URL = database view) and says "look at these"
• A new link was added to the inbox and the user wants it auto-actioned
• User wants to dedupe a Notion link database

What this does

For every row in the Notion DB:

1. Auto-title — if the title is blank or equals the URL:
   • X / Twitter URLs: publish.twitter.com/oembed (no auth needed)
   • Other URLs: scrapes og:title then <title> tag
2. Auto-tag — multi-select tags: ai, coding, tool, aiva, va, marketing, tax, nonprofit, design, personal-finance, event, shopping, startup, meme, x-post, ig-post, github, video, meta-ad
3. Dedupe detection — strips utm_*, fbclid, s=12, etc. and groups by canonical URL
4. Linear suggestions — appends a JSONL row to ~/tools/notion-link-triage/suggestions.jsonl for any entry that hits a high-signal tag (tool, aiva, coding, ai, marketing, tax, nonprofit)

Source database

• DB ID: 1c207f660378818c81a1cfbd2323746b
• URL: https://www.notion.so/1c207f660378818c81a1cfbd2323746b?v=1c207f66037881e990ae000c5c56cbd8
• Schema: Name (title), URL, Tags (multi_select), Date, Created
• Auth: Composio Notion (already connected, see node ~/tools/<your-mcp>/client.cjs status)

Common invocations

# Dry-run preview on first 10 untitled rows (no writes)
python3 ~/tools/notion-link-triage/triage.py --dry-run --limit 10

# Backfill: title + tag every untitled row (default behavior)
python3 ~/tools/notion-link-triage/triage.py

# Process only links added since a date (good for daily incremental)
python3 ~/tools/notion-link-triage/triage.py --since 2026-04-25

# Force-reprocess a single row by page_id
python3 ~/tools/notion-link-triage/triage.py --row <page_id> --force

# Show duplicate groups, no writes
python3 ~/tools/notion-link-triage/triage.py --show-dupes

Workflow: full inbox-zero pass

1. Show current state: python3 ~/tools/notion-link-triage/triage.py --show-dupes
2. Dry-run sanity check on 10 rows → eyeball the output
3. Run the full backfill — takes ~2-3 min for ~280 rows (0.4s sleep between calls)
4. Review ~/tools/notion-link-triage/suggestions.jsonl — each line is a candidate for Linear
5. Group suggestions by tag → propose Linear issues to the user → batch-create via mcp__claude_ai_Linear__save_issue

Workflow: action a single new link

When user pastes a Notion link URL or adds a new entry:

1. Find the page_id (it's the last URL segment after the title slug, with dashes)
2. python3 ~/tools/notion-link-triage/triage.py --row <page_id>
3. Read suggestions.jsonl tail to see the classification
4. If it hit a high-signal tag, propose the Linear issue title + project + priority

Tag → Linear project mapping

Rule: never auto-create Linear issues. Always show the user the proposed title + tags + project, then create only the approved ones.

Daily cron (optional)

If user wants this automated, add a launchd plist on the Mac OR a systemd timer on the VPS that runs the --since "$(date -v-2d +%Y-%m-%d)" pass nightly. Per the user's LLM-vs-Shell rule this is pure shell — use systemd timer on VPS, not openclaw cron. Pattern:

[Unit]
Description=Notion link triage incremental
[Service]
Type=oneshot
ExecStart=/usr/bin/python3 $HOME/notion-link-triage/triage.py --since YYYY-MM-DD
[Install]

Plus a Composio Gmail alert on failure.

Composio actions used

• NOTION_QUERY_DATABASE — paginated fetch of all rows
• NOTION_FETCH_ROW — verify a single row after update
• NOTION_UPDATE_PAGE — write Name (title) + Tags (multi_select) properties

No new OAuth setup — uses the existing Composio Notion connection.

Limitations / known gotchas

• IG posts: Instagram returns a login wall to non-authed clients, so og:title scrape often fails. Title falls back to URL.
• X cookies stale (~/.cookies/x.com): we sidestep this with oEmbed — no auth needed for public tweets.
• Rate limit: 0.4s sleep between rows. ~280 rows → ~2 min. Bump up if oEmbed starts 429-ing (it hasn't yet).
• Duplicate detection is by canonical URL only — doesn't compare titles/content. A tweet quoted in two different X URLs won't merge.
• Tag rules are keyword-based, not semantic. Iterate TAG_RULES in triage.py to improve.

Quick sanity check before reporting done

# Did the title and tags actually persist?
node ~/tools/<your-mcp>/client.cjs execute NOTION_FETCH_ROW '{"page_id":"<id>"}' \
  | grep -E "plain_text|multi_select"

Source files

• ~/tools/notion-link-triage/triage.py — the script
• ~/tools/notion-link-triage/suggestions.jsonl — append-only Linear-issue candidates
• This skill: ~/.claude/skills/notion-link-triage/SKILL.md

Ground-truth gate (MANDATORY)

Before this skill asserts a stakes-bearing fact or takes any outward/irreversible action, apply the global standard — verify against a primary source fetched now, never a cached/remembered value. Full standard: ~/.claude/skills/shared/ground-truth-standard.md.

Verify live before you act or assert (this skill):

• Read the live Notion state before modifying; confirm the target page/block; explicit approval for destructive edits.

Then: dry-run where possible, show the user exactly what will be sent/filed/asserted, get explicit chat approval for any outward action (per CLAUDE.md), capture the confirmation, and write any verified fact back into its source doc. State uncertainty as uncertainty; never assert plausible-but-unverified as fact.